Brain tumors cause serious health problems and brain tumor detection is important for the diagnosis. The detection is a very challenging task due to the complexity in brain structures and in brain tumor patterns. Manual segmentation requires an expertise of highly trained medical specialists and is very time-consuming. Therefore, it’s imperative to develop fully automated brain tumor segmentation systems, i.e., CNN based systems, to accelerate the diagnosis process. The research on developing such systems has been progressed rapidly in recent years. For the systems to be applicable in practice, a good processing quality and reliability are required. Moreover, for a wide range of applications of such systems, a minimization of computation complexity is desirable, which can also result in a minimization of randomness in computation and, consequently, a better performance consistency.
 In this thesis, a new CNN system for brain tumor segmentation is proposed. The CNN in the proposed system is custom-designed with 2 distinguished characters dedicated to optimizing the feature extraction and classification processes. Firstly, there are three paths in its feature extraction block, designed to extract, from the multi-modality input, comprehensive feature information of mono-modality, paired-modality and cross-modality, respectively. Also, it has a particular three-branch classification block to identify the pixels of 4 classes, namely, whole tumor, enhancing tumor, non-enhancing core/necrotic tumor and those in the background. By means of the three branches, a complex multi-class classification problem is decomposed into several simple binary classification problems. Each branch is trained separately so that the parameters are adjusted specifically to suit the detection of one specific kind of tumor areas. The parameters of the convolution layers in the proposed system are determined to suit the specific purposes so that the computation volume for each filtering operations in each layer are just-sufficient, which results in a very simple config of 61,843 parameters in total, while most existing CNN systems require multi-millions.
The performance of the proposed system has been tested extensively with BraTS 2018 and BraTS 2019 data samples. A good mean Dice scores in each experiment has been obtained. The average of the mean Dice scores obtained from ten experiments are very close to each other with very small deviations. In the case of the 10 experiments on BraTS 2018 validation samples, the average Dice scores and their standard deviations are 0.787±0.003, 0.886±0.002, 0.801±0.007, respectively, for enhancing tumor, whole tumor and tumor core. For the validation samples of BraTS 2019 in 10 experiments, the average Dice scores and standard deviations of enhancing tumor, whole tumor and tumor core are 0.751±0.007, 0.885±0.002, 0.776±0.004, respectively. The test results demonstrate that the proposed system is able to perform high-quality segmentation in a consistent manner. Furthermore, it only requires 146G FLOPs to complete a segmentation of the four 3D images (155x240x240x4 voxels) of a single patient case. The extremely low computation complexity of the proposed system will facilitate its implementation/application in various environments.
The high processing quality and low computation complexity of the proposed system make it implementable in various environments. It can be expected that such system will have wide applications in medical image processing.